Debugging Moeum School

Moeum School is Hangman made in Korean.

I uploaded Moeum School to Soosooland in beta status on January 18, 2026.
Simply put, Moeum School is Hangman made in Korean.
It’s a series like Moeum Quiz, Moeum Round, Doori Moeum, and Haroo Moeum .
Currently, only 1st grade Korean language is available for study, and detailed information is in another post.

the foundation of literacy is vocabulary

I heard that the foundation of literacy is vocabulary, so I thought it would be good to teach the meaning of words.
They say when reading, the brain’s processing order is word interpretation → sentence understanding → paragraph understanding → overall text understanding, so if you get stuck in the middle, you can’t move to the next step.
I once read a thick Korean dictionary, just skimming through words and only reading the meanings when I encountered words I didn’t know. It was surprisingly not boring, so I read it twice from beginning to end. If I had known about Ebbinghaus’s forgetting curve when reading the dictionary, I would have committed all the dictionary words to long-term memory, but I didn’t.
I created Moeum School using the principle of Ebbinghaus’s forgetting curve to enhance users’ vocabulary, which is the foundation of literacy.

Moeum School’s Learning System

The words used in Moeum School were extracted from elementary school textbooks. To store word meanings in long-term memory, what you study once is studied again after 3 days, 7 days, 14 days, and 28 days. These are called 1st study, 2nd review, 3rd review, 4th review, and 5th review. Questions answered correctly in the 1st study are considered known and not reviewed.

Bug Discovered!

However, quite some time after uploading Moeum School, a bug was discovered.
I did the 1st study, and when I correctly solved the questions I got wrong in the 1st study during the 2nd review, it gave me a trophy saying I had completed up to the 5th review.
‘Oh, this shouldn’t happen’
Suddenly dopamine rushed as I told Claude we needed to fix the code, and I tried fixing it as Claude instructed.

The Vicious Cycle of Bug Fixing

But then a new bug appeared… I had to explain the game rules again.
Anyway, this couldn’t be verified without loading the word pool stored on GitHub, so I kept uploading to GitHub while testing. Fix one thing and something else breaks, fix that and something else breaks… After doing the 1st review, you should have to wait 7 days but could do it again right away… New bugs kept appearing like that.
In Moeum School, just for 1st grade Korean, there are 49 word pools, and each word pool has about 20 words, so roughly calculating, studying just 1st grade means learning 1000 words. Since the amount of words will become enormous when I make it up to 6th grade and add other subjects, I manage word pools separately and use only one game file.

AI Collaboration(?)

While fixing things with Claude, it said I used up my usage and told me to either pay additional fees according to usage or wait.
While waiting, I went to chat with ChatGPT.
I told it I made a learning game using the principle of Ebbinghaus’s forgetting curve, and it praised me saying that most people don’t even know this system exists, but I directly designed and created a game.
So I asked if the review intervals were appropriate. It said that review intervals doubling each time is an efficient method called spaced repetition.
I said I want to increase reviews more, but since it’s content they’ll study at school anyway, I’m not making it too dense. I mentioned there’s a bug that needs fixing but the file is too long and since it’s ChatGPT’s free plan, I thought it couldn’t read it, but it said code analysis is possible even on the free plan.
After looking at the code, it said a few places needed fixing, so I relayed what ChatGPT said to Claude, and Claude said it was overthinking and that ChatGPT’s opinion was simple and clear (I knew it. Unlike humans, AIs absolutely support each other.) and told me where to fix, but a new bug appeared.

Which Bug is More Serious?

It was time to sleep.I asked Claude.
“The initial bug was giving a trophy when all questions were answered correctly in the 2nd review, and in the 3rd review, questions wrong in the 1st attempt didn’t appear but questions wrong in the 2nd did.
The current bug is that I got 3 wrong in the 1st study but when doing the 2nd review, the number of questions to solve increased to 10-20, and today I did the 3rd review but could do the 4th review today instead of after 14 days, and it recorded that I did up to the 3rd review when I only did the 2nd review. Which is the more serious bug?”
Claude said the current state was more serious and suggested fixing it back to the original state.
“Let’s try again from the beginning in a different window tomorrow. ChatGPT’s opinion might be wrong.”

Claude Opus Deployed!

And today came.
I opened a Claude Opus window and explained the game rules. So Claude made major revisions and newly added the ability to return the trophy and reset when receiving a trophy.
The reason for adding trophy reset is to give users another chance to study when they feel inadequate even after studying up to the 5th review - for example, when they got 10 questions wrong in the 1st study and kept reviewing those 10 questions but couldn’t get a perfect score in the 5th review and feel inadequate.

The Trick for Testing

After making major revisions to Moeum School, it would take 3+7+14+28=52, so 52 days to test. So I asked to change from days to hours so I could check quickly. I was thinking of minutes (3 minutes, 7 minutes, 14 minutes, 28 minutes) in my mind, but Claude changed it to seconds, allowing even faster testing.
Claude came up with the strategic idea of creating a test word pool with only 5 words.
I tested on the iPad I usually play games on, but it wasn’t working properly, then when I tried it on my phone, it worked well.
So thinking the cache was the problem, I cleared the cache storing wrong questions on the iPad and changed test mode from true to false so users could study.​​​​​​​​​​​​​​​​ I changed all the word pool names to prevent bugs caused by cache when users play the game.​​​​​​​​​​​​​​​​

3 Reasons Why the Bug Occurred

There were 3 things that needed to be fixed in the existing code.
1. No mode distinction — reviews were being treated like 1st study
2. Fixed interval calculation — based on 1st study date instead of review completion date
3. Duplicate progress saving — multiple attempts recorded in a single review As expected, Claude Opus seems smarter than Claude Sonnet.
Today, if Claude Opus couldn’t do it properly, I was planning to call in ChatGPT and Gemini to put our heads together and discuss, but in the end, Claude fixed everything.
I wondered if Claude had already fixed it well yesterday but the testing didn’t work properly because of the cache.

🪄 Dev log originally written in Korean | Translated with Claude

Read in Korean 한글로 읽기

Fun Puzzle Playground · Soosooland 
ameDev #wordinterpretation #long-termmemory #Debugging #claude #ChatGPT